Establish prompt.md + schema.json as single source of truth with generator-based CI

[nate-layman]

Problem

Each component currently has multiple hand-authored manifestations (today: prompt.md and skill/SKILL.md; tomorrow: possibly agent/AGENT.md, Gemini extensions, OpenAI Assistants, etc.). Keeping those files consistent by hand drifts — version numbers, descriptions, and prompt bodies all diverge over time. Adding support for a new LLM platform today requires touching N files per component.

Proposal

Declare components/<name>/prompt.md (frontmatter + body) and components/<name>/schema.json the single source of truth for every component. All platform-specific wrappers are derived artifacts produced by a generator script and validated in CI.

Canonical (hand-authored)

• components/<name>/prompt.md — frontmatter (name, version, description, category, domain, status, tags, audience) + prompt body
• components/<name>/schema.json
• components/<name>/README.md, CHANGELOG.md, evals/

Generated (committed, never hand-edited)

• .claude-plugin/marketplace.json at repo root
• components/<name>/.claude-plugin/plugin.json
• components/<name>/skills/<name>/SKILL.md (frontmatter and body both derived from prompt.md)
• future: components/<name>/gemini/gemini-extension.json, components/<name>/openai/assistant.json, etc.

Generator

.github/scripts/generate_wrappers.py scans components/, parses each prompt.md, and writes every derived artifact. Idempotent. Each new platform target is a new render function following the same pattern; canonical sources do not change.

CI

.github/workflows/generate-wrappers.yml runs the generator and fails on git diff --exit-code. Contributors run the script locally before committing — a failure signals that the committed wrappers are stale relative to the canonical sources.

Acceptance criteria

• description: has been migrated from every existing skill/SKILL.md into the corresponding prompt.md frontmatter.
• Legacy components/*/skill/ directories removed.
• Generator script produces marketplace.json, per-component plugin.json, and SKILL.md deterministically.
• CI workflow runs the generator and fails on stale wrappers.
• README documents: canonical vs. generated files, the invariant that wrappers are never hand-edited, and the contributor flow (edit prompt.md → run generator → commit).
• templates/new-component/ reflects the canonical-only layout (no hand-authored skill/ folder).
• Existing .github/scripts/lint_components.py updated to reference the new skills/<name>/SKILL.md path.

Out of scope (tracked separately)

• Additional generator targets for Gemini CLI, OpenAI Assistants, etc. (future issues — this work establishes the primitive they plug into).
• Evaluation runner pairing (Pair components with evaluation-data-sets for reproducibility + accuracy metrics #12).

Relationship to other issues

• Blocks / unblocks Publish prompt-library as a Claude Code plugin marketplace #11 (marketplace publication). The marketplace.json and plugin.json files required by Publish prompt-library as a Claude Code plugin marketplace #11 are outputs of this generator.
• Pair components with evaluation-data-sets for reproducibility + accuracy metrics #12 (evals pairing) will add a metadata.evals field to plugin.json once implemented — that field comes from a canonical source here too.